server/webui: cleanup dual representation approach, simplify to openai-compat

[pwilkin]

Overview

Removes the double-representation from WebUI, bases everything on OpenAI-compat message history

Additional information

Since the current dual representation approach seems to run into a lot of problems (see eg. #21087 ) and will probably run into more when we add the MCP proxy server, I reckoned I'd try to clean it up by removing the double representation and making everything derive from simple chat history - and it seems to be working out pretty well from my tests (tested both normal reasoning and reasoning with tool calling).

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, no way I rewrite all of that by hand 😛

@allozaur @ServeurpersoCom could you take a look and see if it makes sense to you?

[ServeurpersoCom]

This will be a major refactor with many edge cases!

[pwilkin]

@ServeurpersoCom I thought so too, but it went surprisingly smoothly, all tests pass and nothing seems broken at least with cursory testing, so maybe it's not as bad as it looks?

[ServeurpersoCom]

Technically it only touches the storage/rendering side, I'll pull the branch tomorrow and give it a proper test run, because honestly this is how it was originally supposed to be done! This is really the final piece of the MCP client integration, which started as a simple OAI-compat proxy outside the WebUI.

That said, there are a lot of subtle edge cases we've been polishing over the past weeks (streaming interruptions mid-tool-call, partial marker handling, continue generation with open reasoning blocks, tool result truncation, etc.), so I want to make sure nothing regresses before giving the thumbs up!

[ServeurpersoCom]

I'm currently testing it: it's encouraging, it just works. This layer of abstraction was indeed the final blemish on our implementation. And it was much less disruptive than I thought because we naturally converged towards the same structure.

Some obsolete Settings options to remove:

• showRawOutputSwitch
• alwaysShowAgenticTurns
• maxToolPreviewLines

I'm keeping your commit on my everyday server; after a few days of use, I'll be able to spot any edge cases I might have missed.

[allozaur]

I don't thikn showRawOutputSwitch is obsolete, we still might want to use it to debug Markdown rendering

[ServeurpersoCom]

I don't thikn showRawOutputSwitch is obsolete, we still might want to use it to debug Markdown rendering

Ah yes, it needs to be simplified/reworked (broken on the commit for now). We need it for markdown.

[ServeurpersoCom]

Honestly, we can merge quickly and fix what we find, we'll add commits on top and we'll make progress! It works.

[allozaur]

Honestly, we can merge quickly and fix what we find, we'll add commits on top and we'll make progress! It works.

Not exactly 😅 Let's first address any regressions that this PR introduces and then let's merge.

[ServeurpersoCom]

Not exactly 😅 Let's first address any regressions that this PR introduces and then let's merge.

We're going to work together to eradicate them 😅

[pwilkin]

Can you tell me what the regressions found are?

[allozaur]

Can you tell me what the regressions found are?

I am yet to finish reviewing and testing this PR tomorrow, I will let u know from my side.

[allozaur]

@pwilkin i've fixed the agentic turns UI regression + added few more tests. This is a middle-step to what we've internally discussed about having separate structure for chat messages data that is 100% OpenAI compatible, but on the UI end we want to show the assistant + tool messages as a joint agentic chain of messages

[ServeurpersoCom]

Cool, it's definitely more compact this way while retaining the advantages of clean, structured data, good thinking Alek!

[pwilkin]

Looking good!

[ServeurpersoCom]

All that remains is to correct the agent turn display options (visually useful only if there are multiple CoT + Toolcall pairs), and fix the raw Markdown display switch. I've rebased the pull request on my server to try more:)

[allozaur]

All that remains is to correct the agent turn display options (visually useful only if there are multiple CoT + Toolcall pairs), and fix the raw Markdown display switch. I've rebased the pull request on my server to try more:)

these should already be fixed by now

[aldehir]

Good work on this PR. The conversation history looks like it's a proper alternating sequence of assistant/tool roles.

One slight issue, the reasoning content is still not sent back on assistant messages so I opened #21249 to address.